summaryrefslogtreecommitdiff
path: root/net/rds
diff options
context:
space:
mode:
authorLinus Torvalds <torvalds@linux-foundation.org>2009-04-02 21:05:30 -0700
committerLinus Torvalds <torvalds@linux-foundation.org>2009-04-02 21:05:30 -0700
commitef8a97bbc92ec07e3a07a81cc011dc549f8c7a23 (patch)
tree82a95f16d9236bc35a4cfd42ba8cab61981efda8 /net/rds
parent4f032ac4122a77dbabf7a24b2739b2790448180f (diff)
parent6c8ad3b07f7d9efdc41396db6da0aed906922701 (diff)
Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
* git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6: (54 commits) glge: remove unused #include <version.h> dnet: remove unused #include <version.h> tcp: miscounts due to tcp_fragment pcount reset tcp: add helper for counter tweaking due mid-wq change hso: fix for the 'invalid frame length' messages hso: fix for crash when unplugging the device fsl_pq_mdio: Fix compile failure fsl_pq_mdio: Revive UCC MDIO support ucc_geth: Pass proper device to DMA routines, otherwise oops happens i.MX31: Fixing cs89x0 network building to i.MX31ADS tc35815: Fix build error if NAPI enabled hso: add Vendor/Product ID's for new devices ucc_geth: Remove unused header gianfar: Remove unused header kaweth: Fix locking to be SMP-safe net: allow multiple dev per napi with GRO r8169: reset IntrStatus after chip reset ixgbe: Fix potential memory leak/driver panic issue while setting up Tx & Rx ring parameters ixgbe: fix ethtool -A|a behavior ixgbe: Patch to fix driver panic while freeing up tx & rx resources ...
Diffstat (limited to 'net/rds')
-rw-r--r--net/rds/ib.c5
-rw-r--r--net/rds/ib.h28
-rw-r--r--net/rds/ib_cm.c43
-rw-r--r--net/rds/ib_rdma.c43
-rw-r--r--net/rds/ib_recv.c37
-rw-r--r--net/rds/iw.c5
-rw-r--r--net/rds/iw.h28
-rw-r--r--net/rds/iw_cm.c44
-rw-r--r--net/rds/iw_rdma.c44
-rw-r--r--net/rds/iw_recv.c37
-rw-r--r--net/rds/rds.h4
-rw-r--r--net/rds/send.c6
12 files changed, 212 insertions, 112 deletions
diff --git a/net/rds/ib.c b/net/rds/ib.c
index 06a7b798d9a7..4933b380985e 100644
--- a/net/rds/ib.c
+++ b/net/rds/ib.c
@@ -51,6 +51,7 @@ MODULE_PARM_DESC(fmr_message_size, " Max size of a RDMA transfer");
struct list_head rds_ib_devices;
+/* NOTE: if also grabbing ibdev lock, grab this first */
DEFINE_SPINLOCK(ib_nodev_conns_lock);
LIST_HEAD(ib_nodev_conns);
@@ -137,7 +138,7 @@ void rds_ib_remove_one(struct ib_device *device)
kfree(i_ipaddr);
}
- rds_ib_remove_conns(rds_ibdev);
+ rds_ib_destroy_conns(rds_ibdev);
if (rds_ibdev->mr_pool)
rds_ib_destroy_mr_pool(rds_ibdev->mr_pool);
@@ -249,7 +250,7 @@ static int rds_ib_laddr_check(__be32 addr)
void rds_ib_exit(void)
{
rds_info_deregister_func(RDS_INFO_IB_CONNECTIONS, rds_ib_ic_info);
- rds_ib_remove_nodev_conns();
+ rds_ib_destroy_nodev_conns();
ib_unregister_client(&rds_ib_client);
rds_ib_sysctl_exit();
rds_ib_recv_exit();
diff --git a/net/rds/ib.h b/net/rds/ib.h
index 8be563a1363a..069206cae733 100644
--- a/net/rds/ib.h
+++ b/net/rds/ib.h
@@ -108,7 +108,12 @@ struct rds_ib_connection {
/* sending acks */
unsigned long i_ack_flags;
+#ifdef KERNEL_HAS_ATOMIC64
+ atomic64_t i_ack_next; /* next ACK to send */
+#else
+ spinlock_t i_ack_lock; /* protect i_ack_next */
u64 i_ack_next; /* next ACK to send */
+#endif
struct rds_header *i_ack;
struct ib_send_wr i_ack_wr;
struct ib_sge i_ack_sge;
@@ -267,9 +272,17 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn,
/* ib_rdma.c */
int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr);
-int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
-void rds_ib_remove_nodev_conns(void);
-void rds_ib_remove_conns(struct rds_ib_device *rds_ibdev);
+void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
+void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn);
+void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock);
+static inline void rds_ib_destroy_nodev_conns(void)
+{
+ __rds_ib_destroy_conns(&ib_nodev_conns, &ib_nodev_conns_lock);
+}
+static inline void rds_ib_destroy_conns(struct rds_ib_device *rds_ibdev)
+{
+ __rds_ib_destroy_conns(&rds_ibdev->conn_list, &rds_ibdev->spinlock);
+}
struct rds_ib_mr_pool *rds_ib_create_mr_pool(struct rds_ib_device *);
void rds_ib_get_mr_info(struct rds_ib_device *rds_ibdev, struct rds_info_rdma_connection *iinfo);
void rds_ib_destroy_mr_pool(struct rds_ib_mr_pool *);
@@ -355,13 +368,4 @@ rds_ib_data_sge(struct rds_ib_connection *ic, struct ib_sge *sge)
return &sge[1];
}
-static inline void rds_ib_set_64bit(u64 *ptr, u64 val)
-{
-#if BITS_PER_LONG == 64
- *ptr = val;
-#else
- set_64bit(ptr, val);
-#endif
-}
-
#endif
diff --git a/net/rds/ib_cm.c b/net/rds/ib_cm.c
index 0532237bd128..f8e40e1a6038 100644
--- a/net/rds/ib_cm.c
+++ b/net/rds/ib_cm.c
@@ -126,9 +126,7 @@ void rds_ib_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
err = rds_ib_update_ipaddr(rds_ibdev, conn->c_laddr);
if (err)
printk(KERN_ERR "rds_ib_update_ipaddr failed (%d)\n", err);
- err = rds_ib_add_conn(rds_ibdev, conn);
- if (err)
- printk(KERN_ERR "rds_ib_add_conn failed (%d)\n", err);
+ rds_ib_add_conn(rds_ibdev, conn);
/* If the peer gave us the last packet it saw, process this as if
* we had received a regular ACK. */
@@ -616,18 +614,8 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
/*
* Move connection back to the nodev list.
*/
- if (ic->rds_ibdev) {
-
- spin_lock_irq(&ic->rds_ibdev->spinlock);
- BUG_ON(list_empty(&ic->ib_node));
- list_del(&ic->ib_node);
- spin_unlock_irq(&ic->rds_ibdev->spinlock);
-
- spin_lock_irq(&ib_nodev_conns_lock);
- list_add_tail(&ic->ib_node, &ib_nodev_conns);
- spin_unlock_irq(&ib_nodev_conns_lock);
- ic->rds_ibdev = NULL;
- }
+ if (ic->rds_ibdev)
+ rds_ib_remove_conn(ic->rds_ibdev, conn);
ic->i_cm_id = NULL;
ic->i_pd = NULL;
@@ -648,7 +636,11 @@ void rds_ib_conn_shutdown(struct rds_connection *conn)
/* Clear the ACK state */
clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
- rds_ib_set_64bit(&ic->i_ack_next, 0);
+#ifdef KERNEL_HAS_ATOMIC64
+ atomic64_set(&ic->i_ack_next, 0);
+#else
+ ic->i_ack_next = 0;
+#endif
ic->i_ack_recv = 0;
/* Clear flow control state */
@@ -681,6 +673,9 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
INIT_LIST_HEAD(&ic->ib_node);
mutex_init(&ic->i_recv_mutex);
+#ifndef KERNEL_HAS_ATOMIC64
+ spin_lock_init(&ic->i_ack_lock);
+#endif
/*
* rds_ib_conn_shutdown() waits for these to be emptied so they
@@ -701,11 +696,27 @@ int rds_ib_conn_alloc(struct rds_connection *conn, gfp_t gfp)
return 0;
}
+/*
+ * Free a connection. Connection must be shut down and not set for reconnect.
+ */
void rds_ib_conn_free(void *arg)
{
struct rds_ib_connection *ic = arg;
+ spinlock_t *lock_ptr;
+
rdsdebug("ic %p\n", ic);
+
+ /*
+ * Conn is either on a dev's list or on the nodev list.
+ * A race with shutdown() or connect() would cause problems
+ * (since rds_ibdev would change) but that should never happen.
+ */
+ lock_ptr = ic->rds_ibdev ? &ic->rds_ibdev->spinlock : &ib_nodev_conns_lock;
+
+ spin_lock_irq(lock_ptr);
list_del(&ic->ib_node);
+ spin_unlock_irq(lock_ptr);
+
kfree(ic);
}
diff --git a/net/rds/ib_rdma.c b/net/rds/ib_rdma.c
index 69a6289ed672..81033af93020 100644
--- a/net/rds/ib_rdma.c
+++ b/net/rds/ib_rdma.c
@@ -139,7 +139,7 @@ int rds_ib_update_ipaddr(struct rds_ib_device *rds_ibdev, __be32 ipaddr)
return rds_ib_add_ipaddr(rds_ibdev, ipaddr);
}
-int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
+void rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
{
struct rds_ib_connection *ic = conn->c_transport_data;
@@ -148,45 +148,44 @@ int rds_ib_add_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn
BUG_ON(list_empty(&ib_nodev_conns));
BUG_ON(list_empty(&ic->ib_node));
list_del(&ic->ib_node);
- spin_unlock_irq(&ib_nodev_conns_lock);
spin_lock_irq(&rds_ibdev->spinlock);
list_add_tail(&ic->ib_node, &rds_ibdev->conn_list);
spin_unlock_irq(&rds_ibdev->spinlock);
+ spin_unlock_irq(&ib_nodev_conns_lock);
ic->rds_ibdev = rds_ibdev;
-
- return 0;
}
-void rds_ib_remove_nodev_conns(void)
+void rds_ib_remove_conn(struct rds_ib_device *rds_ibdev, struct rds_connection *conn)
{
- struct rds_ib_connection *ic, *_ic;
- LIST_HEAD(tmp_list);
+ struct rds_ib_connection *ic = conn->c_transport_data;
- /* avoid calling conn_destroy with irqs off */
- spin_lock_irq(&ib_nodev_conns_lock);
- list_splice(&ib_nodev_conns, &tmp_list);
- INIT_LIST_HEAD(&ib_nodev_conns);
- spin_unlock_irq(&ib_nodev_conns_lock);
+ /* place conn on nodev_conns_list */
+ spin_lock(&ib_nodev_conns_lock);
- list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) {
- if (ic->conn->c_passive)
- rds_conn_destroy(ic->conn->c_passive);
- rds_conn_destroy(ic->conn);
- }
+ spin_lock_irq(&rds_ibdev->spinlock);
+ BUG_ON(list_empty(&ic->ib_node));
+ list_del(&ic->ib_node);
+ spin_unlock_irq(&rds_ibdev->spinlock);
+
+ list_add_tail(&ic->ib_node, &ib_nodev_conns);
+
+ spin_unlock(&ib_nodev_conns_lock);
+
+ ic->rds_ibdev = NULL;
}
-void rds_ib_remove_conns(struct rds_ib_device *rds_ibdev)
+void __rds_ib_destroy_conns(struct list_head *list, spinlock_t *list_lock)
{
struct rds_ib_connection *ic, *_ic;
LIST_HEAD(tmp_list);
/* avoid calling conn_destroy with irqs off */
- spin_lock_irq(&rds_ibdev->spinlock);
- list_splice(&rds_ibdev->conn_list, &tmp_list);
- INIT_LIST_HEAD(&rds_ibdev->conn_list);
- spin_unlock_irq(&rds_ibdev->spinlock);
+ spin_lock_irq(list_lock);
+ list_splice(list, &tmp_list);
+ INIT_LIST_HEAD(list);
+ spin_unlock_irq(list_lock);
list_for_each_entry_safe(ic, _ic, &tmp_list, ib_node) {
if (ic->conn->c_passive)
diff --git a/net/rds/ib_recv.c b/net/rds/ib_recv.c
index 5061b5502162..36d931573ff4 100644
--- a/net/rds/ib_recv.c
+++ b/net/rds/ib_recv.c
@@ -395,10 +395,37 @@ void rds_ib_recv_init_ack(struct rds_ib_connection *ic)
* room for it beyond the ring size. Send completion notices its special
* wr_id and avoids working with the ring in that case.
*/
+#ifndef KERNEL_HAS_ATOMIC64
static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
int ack_required)
{
- rds_ib_set_64bit(&ic->i_ack_next, seq);
+ unsigned long flags;
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ ic->i_ack_next = seq;
+ if (ack_required)
+ set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+}
+
+static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
+{
+ unsigned long flags;
+ u64 seq;
+
+ clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ seq = ic->i_ack_next;
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+
+ return seq;
+}
+#else
+static void rds_ib_set_ack(struct rds_ib_connection *ic, u64 seq,
+ int ack_required)
+{
+ atomic64_set(&ic->i_ack_next, seq);
if (ack_required) {
smp_mb__before_clear_bit();
set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
@@ -410,8 +437,10 @@ static u64 rds_ib_get_ack(struct rds_ib_connection *ic)
clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
smp_mb__after_clear_bit();
- return ic->i_ack_next;
+ return atomic64_read(&ic->i_ack_next);
}
+#endif
+
static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credits)
{
@@ -464,6 +493,10 @@ static void rds_ib_send_ack(struct rds_ib_connection *ic, unsigned int adv_credi
* - i_ack_next, which is the last sequence number we received
*
* Potentially, send queue and receive queue handlers can run concurrently.
+ * It would be nice to not have to use a spinlock to synchronize things,
+ * but the one problem that rules this out is that 64bit updates are
+ * not atomic on all platforms. Things would be a lot simpler if
+ * we had atomic64 or maybe cmpxchg64 everywhere.
*
* Reconnecting complicates this picture just slightly. When we
* reconnect, we may be seeing duplicate packets. The peer
diff --git a/net/rds/iw.c b/net/rds/iw.c
index 1b56905c4c08..b732efb5b634 100644
--- a/net/rds/iw.c
+++ b/net/rds/iw.c
@@ -51,6 +51,7 @@ MODULE_PARM_DESC(fastreg_message_size, " Max size of a RDMA transfer (fastreg MR
struct list_head rds_iw_devices;
+/* NOTE: if also grabbing iwdev lock, grab this first */
DEFINE_SPINLOCK(iw_nodev_conns_lock);
LIST_HEAD(iw_nodev_conns);
@@ -145,7 +146,7 @@ void rds_iw_remove_one(struct ib_device *device)
}
spin_unlock_irq(&rds_iwdev->spinlock);
- rds_iw_remove_conns(rds_iwdev);
+ rds_iw_destroy_conns(rds_iwdev);
if (rds_iwdev->mr_pool)
rds_iw_destroy_mr_pool(rds_iwdev->mr_pool);
@@ -258,7 +259,7 @@ static int rds_iw_laddr_check(__be32 addr)
void rds_iw_exit(void)
{
rds_info_deregister_func(RDS_INFO_IWARP_CONNECTIONS, rds_iw_ic_info);
- rds_iw_remove_nodev_conns();
+ rds_iw_destroy_nodev_conns();
ib_unregister_client(&rds_iw_client);
rds_iw_sysctl_exit();
rds_iw_recv_exit();
diff --git a/net/rds/iw.h b/net/rds/iw.h
index 0ddda34f2a1c..b4fb27252895 100644
--- a/net/rds/iw.h
+++ b/net/rds/iw.h
@@ -131,7 +131,12 @@ struct rds_iw_connection {
/* sending acks */
unsigned long i_ack_flags;
+#ifdef KERNEL_HAS_ATOMIC64
+ atomic64_t i_ack_next; /* next ACK to send */
+#else
+ spinlock_t i_ack_lock; /* protect i_ack_next */
u64 i_ack_next; /* next ACK to send */
+#endif
struct rds_header *i_ack;
struct ib_send_wr i_ack_wr;
struct ib_sge i_ack_sge;
@@ -294,9 +299,17 @@ void rds_iw_cm_connect_complete(struct rds_connection *conn,
/* ib_rdma.c */
int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_id);
-int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
-void rds_iw_remove_nodev_conns(void);
-void rds_iw_remove_conns(struct rds_iw_device *rds_iwdev);
+void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
+void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn);
+void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock);
+static inline void rds_iw_destroy_nodev_conns(void)
+{
+ __rds_iw_destroy_conns(&iw_nodev_conns, &iw_nodev_conns_lock);
+}
+static inline void rds_iw_destroy_conns(struct rds_iw_device *rds_iwdev)
+{
+ __rds_iw_destroy_conns(&rds_iwdev->conn_list, &rds_iwdev->spinlock);
+}
struct rds_iw_mr_pool *rds_iw_create_mr_pool(struct rds_iw_device *);
void rds_iw_get_mr_info(struct rds_iw_device *rds_iwdev, struct rds_info_rdma_connection *iinfo);
void rds_iw_destroy_mr_pool(struct rds_iw_mr_pool *);
@@ -383,13 +396,4 @@ rds_iw_data_sge(struct rds_iw_connection *ic, struct ib_sge *sge)
return &sge[1];
}
-static inline void rds_iw_set_64bit(u64 *ptr, u64 val)
-{
-#if BITS_PER_LONG == 64
- *ptr = val;
-#else
- set_64bit(ptr, val);
-#endif
-}
-
#endif
diff --git a/net/rds/iw_cm.c b/net/rds/iw_cm.c
index 57ecb3d4b8a5..a416b0d492b1 100644
--- a/net/rds/iw_cm.c
+++ b/net/rds/iw_cm.c
@@ -86,9 +86,7 @@ void rds_iw_cm_connect_complete(struct rds_connection *conn, struct rdma_cm_even
err = rds_iw_update_cm_id(rds_iwdev, ic->i_cm_id);
if (err)
printk(KERN_ERR "rds_iw_update_ipaddr failed (%d)\n", err);
- err = rds_iw_add_conn(rds_iwdev, conn);
- if (err)
- printk(KERN_ERR "rds_iw_add_conn failed (%d)\n", err);
+ rds_iw_add_conn(rds_iwdev, conn);
/* If the peer gave us the last packet it saw, process this as if
* we had received a regular ACK. */
@@ -637,19 +635,8 @@ void rds_iw_conn_shutdown(struct rds_connection *conn)
* Move connection back to the nodev list.
* Remove cm_id from the device cm_id list.
*/
- if (ic->rds_iwdev) {
-
- spin_lock_irq(&ic->rds_iwdev->spinlock);
- BUG_ON(list_empty(&ic->iw_node));
- list_del(&ic->iw_node);
- spin_unlock_irq(&ic->rds_iwdev->spinlock);
-
- spin_lock_irq(&iw_nodev_conns_lock);
- list_add_tail(&ic->iw_node, &iw_nodev_conns);
- spin_unlock_irq(&iw_nodev_conns_lock);
- rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
- ic->rds_iwdev = NULL;
- }
+ if (ic->rds_iwdev)
+ rds_iw_remove_conn(ic->rds_iwdev, conn);
rdma_destroy_id(ic->i_cm_id);
@@ -672,7 +659,11 @@ void rds_iw_conn_shutdown(struct rds_connection *conn)
/* Clear the ACK state */
clear_bit(IB_ACK_IN_FLIGHT, &ic->i_ack_flags);
- rds_iw_set_64bit(&ic->i_ack_next, 0);
+#ifdef KERNEL_HAS_ATOMIC64
+ atomic64_set(&ic->i_ack_next, 0);
+#else
+ ic->i_ack_next = 0;
+#endif
ic->i_ack_recv = 0;
/* Clear flow control state */
@@ -706,6 +697,9 @@ int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
INIT_LIST_HEAD(&ic->iw_node);
mutex_init(&ic->i_recv_mutex);
+#ifndef KERNEL_HAS_ATOMIC64
+ spin_lock_init(&ic->i_ack_lock);
+#endif
/*
* rds_iw_conn_shutdown() waits for these to be emptied so they
@@ -726,11 +720,27 @@ int rds_iw_conn_alloc(struct rds_connection *conn, gfp_t gfp)
return 0;
}
+/*
+ * Free a connection. Connection must be shut down and not set for reconnect.
+ */
void rds_iw_conn_free(void *arg)
{
struct rds_iw_connection *ic = arg;
+ spinlock_t *lock_ptr;
+
rdsdebug("ic %p\n", ic);
+
+ /*
+ * Conn is either on a dev's list or on the nodev list.
+ * A race with shutdown() or connect() would cause problems
+ * (since rds_iwdev would change) but that should never happen.
+ */
+ lock_ptr = ic->rds_iwdev ? &ic->rds_iwdev->spinlock : &iw_nodev_conns_lock;
+
+ spin_lock_irq(lock_ptr);
list_del(&ic->iw_node);
+ spin_unlock_irq(lock_ptr);
+
kfree(ic);
}
diff --git a/net/rds/iw_rdma.c b/net/rds/iw_rdma.c
index 1c02a8f952d0..dcdb37da80f2 100644
--- a/net/rds/iw_rdma.c
+++ b/net/rds/iw_rdma.c
@@ -196,7 +196,7 @@ int rds_iw_update_cm_id(struct rds_iw_device *rds_iwdev, struct rdma_cm_id *cm_i
return rds_iw_add_cm_id(rds_iwdev, cm_id);
}
-int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
+void rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
{
struct rds_iw_connection *ic = conn->c_transport_data;
@@ -205,45 +205,45 @@ int rds_iw_add_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn
BUG_ON(list_empty(&iw_nodev_conns));
BUG_ON(list_empty(&ic->iw_node));
list_del(&ic->iw_node);
- spin_unlock_irq(&iw_nodev_conns_lock);
spin_lock_irq(&rds_iwdev->spinlock);
list_add_tail(&ic->iw_node, &rds_iwdev->conn_list);
spin_unlock_irq(&rds_iwdev->spinlock);
+ spin_unlock_irq(&iw_nodev_conns_lock);
ic->rds_iwdev = rds_iwdev;
-
- return 0;
}
-void rds_iw_remove_nodev_conns(void)
+void rds_iw_remove_conn(struct rds_iw_device *rds_iwdev, struct rds_connection *conn)
{
- struct rds_iw_connection *ic, *_ic;
- LIST_HEAD(tmp_list);
+ struct rds_iw_connection *ic = conn->c_transport_data;
- /* avoid calling conn_destroy with irqs off */
- spin_lock_irq(&iw_nodev_conns_lock);
- list_splice(&iw_nodev_conns, &tmp_list);
- INIT_LIST_HEAD(&iw_nodev_conns);
- spin_unlock_irq(&iw_nodev_conns_lock);
+ /* place conn on nodev_conns_list */
+ spin_lock(&iw_nodev_conns_lock);
- list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node) {
- if (ic->conn->c_passive)
- rds_conn_destroy(ic->conn->c_passive);
- rds_conn_destroy(ic->conn);
- }
+ spin_lock_irq(&rds_iwdev->spinlock);
+ BUG_ON(list_empty(&ic->iw_node));
+ list_del(&ic->iw_node);
+ spin_unlock_irq(&rds_iwdev->spinlock);
+
+ list_add_tail(&ic->iw_node, &iw_nodev_conns);
+
+ spin_unlock(&iw_nodev_conns_lock);
+
+ rds_iw_remove_cm_id(ic->rds_iwdev, ic->i_cm_id);
+ ic->rds_iwdev = NULL;
}
-void rds_iw_remove_conns(struct rds_iw_device *rds_iwdev)
+void __rds_iw_destroy_conns(struct list_head *list, spinlock_t *list_lock)
{
struct rds_iw_connection *ic, *_ic;
LIST_HEAD(tmp_list);
/* avoid calling conn_destroy with irqs off */
- spin_lock_irq(&rds_iwdev->spinlock);
- list_splice(&rds_iwdev->conn_list, &tmp_list);
- INIT_LIST_HEAD(&rds_iwdev->conn_list);
- spin_unlock_irq(&rds_iwdev->spinlock);
+ spin_lock_irq(list_lock);
+ list_splice(list, &tmp_list);
+ INIT_LIST_HEAD(list);
+ spin_unlock_irq(list_lock);
list_for_each_entry_safe(ic, _ic, &tmp_list, iw_node) {
if (ic->conn->c_passive)
diff --git a/net/rds/iw_recv.c b/net/rds/iw_recv.c
index a1931f0027a2..fde470fa50d5 100644
--- a/net/rds/iw_recv.c
+++ b/net/rds/iw_recv.c
@@ -395,10 +395,37 @@ void rds_iw_recv_init_ack(struct rds_iw_connection *ic)
* room for it beyond the ring size. Send completion notices its special
* wr_id and avoids working with the ring in that case.
*/
+#ifndef KERNEL_HAS_ATOMIC64
static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq,
int ack_required)
{
- rds_iw_set_64bit(&ic->i_ack_next, seq);
+ unsigned long flags;
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ ic->i_ack_next = seq;
+ if (ack_required)
+ set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+}
+
+static u64 rds_iw_get_ack(struct rds_iw_connection *ic)
+{
+ unsigned long flags;
+ u64 seq;
+
+ clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
+
+ spin_lock_irqsave(&ic->i_ack_lock, flags);
+ seq = ic->i_ack_next;
+ spin_unlock_irqrestore(&ic->i_ack_lock, flags);
+
+ return seq;
+}
+#else
+static void rds_iw_set_ack(struct rds_iw_connection *ic, u64 seq,
+ int ack_required)
+{
+ atomic64_set(&ic->i_ack_next, seq);
if (ack_required) {
smp_mb__before_clear_bit();
set_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
@@ -410,8 +437,10 @@ static u64 rds_iw_get_ack(struct rds_iw_connection *ic)
clear_bit(IB_ACK_REQUESTED, &ic->i_ack_flags);
smp_mb__after_clear_bit();
- return ic->i_ack_next;
+ return atomic64_read(&ic->i_ack_next);
}
+#endif
+
static void rds_iw_send_ack(struct rds_iw_connection *ic, unsigned int adv_credits)
{
@@ -464,6 +493,10 @@ static void rds_iw_send_ack(struct rds_iw_connection *ic, unsigned int adv_credi
* - i_ack_next, which is the last sequence number we received
*
* Potentially, send queue and receive queue handlers can run concurrently.
+ * It would be nice to not have to use a spinlock to synchronize things,
+ * but the one problem that rules this out is that 64bit updates are
+ * not atomic on all platforms. Things would be a lot simpler if
+ * we had atomic64 or maybe cmpxchg64 everywhere.
*
* Reconnecting complicates this picture just slightly. When we
* reconnect, we may be seeing duplicate packets. The peer
diff --git a/net/rds/rds.h b/net/rds/rds.h
index 060400704979..619f0a30a4e5 100644
--- a/net/rds/rds.h
+++ b/net/rds/rds.h
@@ -28,6 +28,10 @@
*/
#define RDS_PORT 18634
+#ifdef ATOMIC64_INIT
+#define KERNEL_HAS_ATOMIC64
+#endif
+
#ifdef DEBUG
#define rdsdebug(fmt, args...) pr_debug("%s(): " fmt, __func__ , ##args)
#else
diff --git a/net/rds/send.c b/net/rds/send.c
index 1b37364656f0..104fe033203d 100644
--- a/net/rds/send.c
+++ b/net/rds/send.c
@@ -615,7 +615,7 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
{
struct rds_message *rm, *tmp;
struct rds_connection *conn;
- unsigned long flags;
+ unsigned long flags, flags2;
LIST_HEAD(list);
int wake = 0;
@@ -651,9 +651,9 @@ void rds_send_drop_to(struct rds_sock *rs, struct sockaddr_in *dest)
list_for_each_entry(rm, &list, m_sock_item) {
/* We do this here rather than in the loop above, so that
* we don't have to nest m_rs_lock under rs->rs_lock */
- spin_lock(&rm->m_rs_lock);
+ spin_lock_irqsave(&rm->m_rs_lock, flags2);
rm->m_rs = NULL;
- spin_unlock(&rm->m_rs_lock);
+ spin_unlock_irqrestore(&rm->m_rs_lock, flags2);
/*
* If we see this flag cleared then we're *sure* that someone